Indirect Encoding for Mapping Parameter Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training mappings, such as neural networks, face inefficiencies due to the exponential growth of independently tunable parameters with the number of dimensions, leading to impractical training times and high memory requirements, especially with large data sets and complex mappings.
Innovation Solution
The use of indirect encodings, where mapping parameters are determined based on relationship information such as location, geometry, and task-specific data, allowing for efficient optimization by propagating error information from the mapping to the indirect encoding, reducing the number of independently adjustable parameters and enabling faster training with less data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional direct encoding methods are used to represent mapping parameters, then the mapping can be trained with standard optimization algorithms, but the number of independently tunable parameters grows exponentially with the number of dimensions, leading to impractical training times and high memory requirements
Solution Approach 1:
The patent transforms the parameter space from direct encoding to indirect encoding by introducing a new dimension. Instead of representing parameters w directly, the system uses an indirect encoding function g(z,v) where z represents geometric or structural relationships and v are latent variables. This dimensional transformation reduces the effective number of parameters while maintaining representational capacity, directly addressing the exponential parameter growth problem.
Solution Approach 2:
The indirect encoding function g serves as an intermediary between the latent variables v and the actual mapping parameters w. Rather than optimizing w directly, the system optimizes v and lets g transform it into the parameter space. This intermediary layer enables efficient optimization by propagating error information through the encoding function, avoiding the need to handle exponentially many parameters directly.
2Loss of time
If the number of mapping parameters is reduced through indirect encoding, then training time and memory requirements decrease, but the complexity of determining parameter information from relationship information increases
Solution Approach 1:
The patent segments the parameter determination process into distinct components: relationship information extraction (geometric/structural relationships z), latent variable optimization (v), and parameter generation through the encoding function (g). This segmentation allows each component to be optimized independently, reducing overall complexity despite the indirect nature of parameter determination.
Solution Approach 2:
The system implements feedback loops where error information from the mapping output flows back through the indirect encoding function to adjust the latent variables v. This feedback mechanism automates the parameter determination process, reducing manual complexity while enabling efficient optimization of the indirect encoding to minimize error metrics.
Data Source
AI summary
Methods, systems and apparatuses, including computer programs encoded on computer storage media, are provided for learning or optimizing an indirect encoding of a mapping from digitally-encoded input arrays to digitally-encoded output arrays, with numerous technical advantages in terms of efficiency and effectiveness.